Extremely low running time when profiling transpiled muGraphs

[wmdi]

When profiling transpiled muGrpahs, some results are extremely low and are close to kernel launch time. For example, in the gated_mlp example, some muGraphs only consume ~0.004ms in the catalyst cluster. This is an indication that we may have kernel launch error in the generated cuda programs.

[jiazhihao]

I suspect this is because some of the generated CUDA kernels cannot be successfully executed on A5000 GPUs. One possibility is that the required smem size exceeds the hardware limit. @xinhaoc Can you take a look at the gated_mlp issue on A5000?